Black Forest Labs Unveils FLUX 3 AI: Ditches Stills for Video—And Robotic Palms

Black Forest Labs launched FLUX 3 on Thursday, and for the primary time, the corporate’s flagship mannequin generates video as a substitute of simply nonetheless pictures. The German AI lab, recognized for the FLUX line of picture turbines, educated the brand new system on pictures, video, and audio directly, inside one shared system.

That’s what is named multimodality: one mannequin studying a number of varieties of data collectively as a substitute of separate instruments bolted aspect by aspect.

The video aspect is the headline function. FLUX 3 produces clips as much as 20 seconds lengthy, with audio generated alongside the image and synced to what’s occurring on display screen—dialogue, sound results, ambient noise. In early evaluations, human reviewers most well-liked FLUX 3’s output over Runway Gen-4.5 in 77% of head-to-head comparisons and over Luma Ray 3.2 in 93%. It appears to be barely higher than Gemini Omni and Seedance, beating these fashions in 52% of the evaluations.

After all, that’s a desire take a look at, not a hard and fast scoring rubric: evaluators merely watch two clips and decide the one that appears and sounds extra convincing, and BFL counts how typically FLUX 3 wins.

Apart from that, the mannequin appears to be very competent on nonetheless pictures too, following its legacy. BFL shared a number of pictures, and FLUX 3 appears to be very versatile and able to producing a broad number of kinds past photorealism.

BFL frames this as greater than a content material instrument. “A mannequin that solely learns pictures can solely generate pictures,” mentioned co-founder and CEO Robin Rombach. The corporate’s wager is that studying to foretell video additionally means studying the physics beneath it—weight, contact, timing—which is strictly what a machine wants to maneuver by the bodily world.

That wager has a reputation: FLUX-mimic. Constructed with Zurich-based mimic robotics, it takes FLUX 3’s video-prediction engine and provides a light-weight “decoder”—a small add-on part that interprets the mannequin’s inside sense of how issues transfer into precise robotic motions. Automotive maker Audi is already testing it on duties like becoming versatile door seals, work that standard automation has struggled to deal with.

“Audi represents the type of manufacturing accomplice we constructed FLUX-mimic for,” mentioned mimic co-founder Stephan-Daniel Gravert. Audi’s Christoph Schneider mentioned the robots now “resolve advanced soft-body manipulation work” that older machines couldn’t contact. BFL says the complete system reacts in about 101 milliseconds, within the neighborhood of human visible reflexes.

FLUX’s rise didn’t occur in a vacuum. Based in August 2024 by veteran researchers who’d helped construct the unique Secure Diffusion fashions at Stability AI, Black Forest Labs launched Flux fashions that beat MidJourney and outclassed Stability’s personal underwhelming Stable Diffusion 3.

The open-source Flux Dev and Schnell fashions grabbed the “finest open supply picture generator” title that AI artists had anticipated Secure Diffusion 3.5, Stability’s do-over, to ultimately reclaim.

It by no means did. FLUX 1.1 Professional went on to high the Artificial Analysis image arena that October. That one wasn’t open supply, although.

BFL launched FLUX.2 in November 2025 nevertheless it wasn’t as common. The open-source crown held by the unique Flux lasted till Alibaba’s Z-Image Turbo dethroned it in late 2025, matching its high quality on decrease finish client graphics playing cards. “That is what SD3 was imagined to be,” one CivitAI person wrote on the time.

FLUX 3 is BFL’s comeback, and it isn’t absolutely open but. Video and Motion are in early entry now by APIs and choose companions, mimic robotics amongst them, with picture technology following “within the coming weeks,” per BFL. The open-weight Dev model, the one tier BFL plans to launch for native use, isn’t due till later in 2026.